xcw520/dsh-dania-voice

Local GPT-SoVITS text-to-speech for DeepSeek Harness: reads assistant replies aloud in a fine-tuned voice from a draggable mascot, with no text or audio leaving the machine.

The plugin hooks the session event stream, takes the final assistant reply, strips code blocks, tables and links, then merges the rest into segments of at most 120 characters and posts them to a local GPT-SoVITS server. Generated wav files are cached and queued, and the browser polls for the next clip and plays it through Web Audio, which keeps audio on the client instead of the headless host. A small mascot hanging at the top of the interface exposes the controls so reading aloud can be toggled per session without a settings page.

Vision & Multimodal ★ 4
View on GitHub ↗

Install

dsh plugin --profile web add dsh-dania-voice

Copied verbatim from the README's Install block. package dsh-dania-voice, registry latest 1.0.0, and its repository field resolves back to github.com/xcw520/dsh-dania-voice. After installing, open the mascot panel and run its one-click setup, which creates a Python venv and installs PyTorch plus the base and voice models; text-to-speech needs a GPT-SoVITS api_v2.py server listening on port 9880.

Compatibility

Requires DSH >= 0.1.7-rc.1 and Python 3.10-3.12 on the host, which the installer looks for but does not install. An NVIDIA GPU with 8 GB or more VRAM is recommended and the CPU path is several times slower; about 5 GB of disk is needed for PyTorch, the base models and the voice model. Playback runs in the browser because the host process has no audio device, so the voice comes out of the machine you are sitting at.

Details

Alternatives

wads-dev/dsh-live-voice · kittcat-lab/dsh-kitt-voice · PiyotaHu/muxiva-dsh-voice

More plugins in Vision & Multimodal

Browse more in Vision & Multimodal

Guides for Vision & Multimodal plugins